Why I use AI to simulate the customer before I write a single word

When I explore new solutions or vendors, most B2B SaaS websites are full of statements that read something like this: “Transform your workflow with our revolutionary solution”. It’s the kind of sentence that feels smart at 2 am, fuelled by caffeine and optimism. Put it in front of a real focus group, though, and the feedback would be ruthless.

The genuinely good news is that AI changes this dynamic. You can have something very close to a focus group, even at 2 am. By building a synthetic customer persona and running your copy past it before you ship, you introduce friction where most teams rely on vibes.

Thanks for reading Cowritten.ai! Subscribe for free to receive new posts and support my work.

I’ve been experimenting with synthetic customers for a few months and, if done right, it’s more effective than you might think.

This shift from using AI as a writing assistant to working with it as a rehearsal partner has fundamentally changed how I approach copywriting.

I don’t just use AI to generate the copy. I make it argue with it. Let it poke holes in it. Let it tell me exactly where real buyers will tune out before I spend a penny on ads or commit resources to a campaign that might fall flat.

The problem with validation latency

Every content team knows the sinking feeling: you’ve launched the campaign, committed the budget, and then the data arrives. The messaging isn’t landing. The positioning feels off. Customers are confused, or worse, they’re simply not responding. By the time you understand why, you’ve already burned through resources and market opportunity.

Traditional market research methods create a fundamental timing problem. Organising a focus group takes weeks of recruitment and scheduling. A proper survey requires participant compensation, management overhead, and analysis time that can stretch into months. For niche B2B audiences (procurement managers, revenue executives, decision-makers), the recruitment challenge becomes even more daunting. The timeline is long, the cost is high, and by the time insights arrive, market conditions may have already shifted.

This validation latency creates a cascade of compromises. Teams either skip testing entirely (risking expensive failures) or do minimal testing due to resource constraints. Neither option is sustainable when customer expectations keep rising and launch windows keep shrinking.

Synthetic personas are rehearsal rooms, not oracles

What a synthetic persona is (and what it is not)

Think of synthetic personas as rehearsal partners, not fortune tellers. They’re dynamic, interactive models trained on real research data that can respond to prompts in ways that mirror how actual members of your target segment might think and speak. When you show them marketing copy, a cost-conscious small business owner persona might raise objections around price and implementation burden. A sustainability-focused millennial might zero in on brand values and environmental claims.

These aren’t chatbots following scripts. Modern synthetic personas leverage the vast amounts of human discourse that large language models have absorbed (customer reviews, social media conversations, support tickets, survey responses). They generate responses that reflect learned patterns in how people with specific characteristics typically evaluate products and communicate their concerns.

The sophistication has reached remarkable levels. Researchers at Stanford and Google DeepMind created AI agents that simulated over 1,000 individuals’ personalities with 85% accuracy on standard survey instruments. Even more compelling, PyMC Labs and Colgate-Palmolive demonstrated that LLMs can reproduce human purchase intent with 90% accuracy using their Semantic Similarity Rating technique.

But here’s what synthetic personas are not: they’re not replacements for real customer understanding. They’re accelerants. They help you explore possibilities, raise questions worth investigating, and identify assumptions that might be wrong. They excel at hypothesis generation and early exploration but struggle with capturing genuine lived experience or deep cultural context.

Three data sources that make a persona honest (sales transcripts, negative reviews, support tickets)

The quality of your synthetic personas depends entirely on the data that shapes them. Generic demographic information produces generic responses. But when you build personas from the messy reality of actual customer interactions, you get something far more valuable.

Start with your sales transcripts, particularly from lost deals. These conversations reveal the exact language prospects use when they’re not convinced. They show you the objections that actually kill deals, not the ones your sales team assumes are important. Feed these transcripts into your persona construction, and your synthetic customers will mirror those real hesitations.

Negative reviews are gold. They’re unfiltered expressions of disappointment, frustration, and unmet expectations. When customers write one-star reviews, they’re telling you exactly where your messaging created false expectations or your product failed to deliver. This data teaches your personas to spot the same disconnects in new copy before it goes live.

Support tickets complete the picture. They show you where confusion lives, what language causes misunderstanding, and which promises feel hollow once customers actually engage with your product. The most effective personas combine all three sources, creating multi-dimensional models that reflect not just demographics but actual behavioural patterns and communication styles.

The simulation workflow: a practical framework

Step 1: collect and ground the persona (data checklist)

Building a useful synthetic persona starts with data collection. You need:

Sales call extracts from lost deals: Pull 20-30 transcript excerpts where prospects explained why they didn’t buy

Negative reviews: Collect your worst feedback, organised by complaint theme

Support tickets: Export common issues, particularly those revealing confusion about your messaging

Buyer journey notes: Document where prospects typically drop off

Competitive objections: List what prospects say when choosing competitors

Organise this data into a persona brief with these fields:

Role and responsibilities

Key pain points (in their words, not yours)

Budget constraints and approval process

Typical language patterns and jargon

Common objections and concerns

Success metrics and KPIs they care about

Step 2: switch roles, brief the model to be a real buyer (prompt pattern)

Here’s the prompt pattern that transforms an LLM into your synthetic customer:

You are [specific role] at [company type] with [context].Your background:- [Insert 3-4 key characteristics from persona brief]- You’ve been researching [solution category] for [timeframe]- Your main concerns are: [list from data]Your communication style:- [Describe how they speak, based on transcripts]- You’re particularly sensitive to: [triggers from negative reviews]- You immediately distrust: [patterns from lost deals]I’m going to show you some marketing copy. Read it as if you’re evaluating whether to engage further. Be brutally honest about what makes you want to leave the page or delete the email.For each piece of copy, provide:1. Confidence score (0-100): How likely you are to continue engaging2. Delete trigger: The specific phrase that would make you stop reading3. The gap: What you expected vs what you got

Step 3: run the confidence score gate (scoring template and rules)

The confidence score becomes your quality gate. Here’s how to use it:

Apply this rule strictly: nothing below 80 goes live. It’s not about perfection here, but rather about avoiding the obvious failures that waste budget and opportunity.

When the persona gives you a score, always ask: “What single change would increase this score by 10 points?” The answer often reveals the specific edit that transforms mediocre copy into something that connects.

Step 4: turn feedback into edits and iterate (triage framework)

Not all feedback is equal. Here’s how to triage what you get:

Act immediately on:

Delete triggers (phrases that stop engagement)

Confusion points (where meaning is unclear)

Trust breakers (claims that feel exaggerated)

Test and refine:

Alternative phrasings the persona suggests

Structural changes (reordering benefits)

Tone adjustments

Consider but verify:

Feature requests or product feedback

Pricing concerns

Competitive comparisons

Run your edited copy through the same persona again. The score should improve. If it doesn’t, you’ve either misunderstood the feedback or uncovered a deeper disconnect that needs addressing.

Concrete examples and smell tests

Headline and email subject tests (before/after)

Here’s what transformation looks like in practice:

Before: “Revolutionary AI platform for modern teams”

Confidence score: 45

Delete trigger: “Revolutionary”

The gap: “Expected specific capability, got marketing fluff”

After: “Cut content production time by 60% without losing your brand voice”

Confidence score: 85

Positive signal: “Specific metric + addresses real concern”

Remaining gap: “Want to know how it maintains voice”

Before: “Transform your content workflow”

Confidence score: 35

Delete trigger: “Transform” (overused, vague)

The gap: “Sounds like every other vendor email”

After: “How Acme Inc. writes 50 help articles per week with 3 people”

Confidence score: 88

Positive signal: “Specific example, credible scale”

Follow-up: “Now curious about the method”

Landing page flow test (where buyers drop off)

Feed your landing page copy section by section to the persona. Ask after each: “Would you keep scrolling? Why or why not?”

Common drop-off points personas identify:

Hero sections that don’t state the specific problem solved

Feature lists without connecting to outcomes

Social proof that feels generic (”Join 10,000+ companies”)

CTAs that ask for commitment before showing value

One pattern I see repeatedly: personas bail when copy shifts from addressing their problem to talking about the product’s technical architecture. They want to know what changes for them, not how your backend works.

Limits, ethics and calibration

Where synthetic personas mislead you (low variability, cultural blind spots)

Synthetic personas have real limitations you must account for. Research shows they tend to cluster around average responses, showing less diversity than real humans. They might miss the edge cases (the passionate advocates or fierce critics who often provide the most valuable insights).

Cultural blind spots are particularly problematic. Personas trained primarily on English-language, Western consumer data struggle to represent other markets authentically. They might miss culture-specific purchase drivers or misunderstand regional variations in how customers evaluate solutions.

There’s also what researchers call “hyper-accuracy distortion.” Synthetic personas sound unnaturally confident and coherent. Real people express uncertainty, contradiction, and ambivalence. A synthetic persona might give you perfectly logical reasoning for a preference, while a real person might just say “it feels off” without being able to articulate why.

These aren’t reasons to abandon synthetic testing. They’re reasons to use it appropriately: for early exploration, hypothesis generation, and catching obvious problems. Not for final validation or high-stakes decisions requiring genuine human insight.

How to validate and build trust

Building confidence in synthetic research requires validation loops. Run synthetic tests, make decisions based on them, then verify with small real-user samples. Track how often synthetic feedback aligns with actual customer response.

Key calibration metrics to monitor:

Correlation between synthetic confidence scores and real conversion rates

Accuracy of predicted objections vs actual sales conversations

Alignment between synthetic preferences and A/B test results

Start with low-stakes tests. Use synthetic personas to choose email subject lines or headline variations, then measure actual open rates and engagement. As correlation patterns emerge, you can gradually increase reliance on synthetic insights for bigger decisions.

Document everything. Which synthetic insights proved accurate? Which missed the mark? Over time, you’ll develop calibrated intuition about when to trust the simulation and when to demand human validation.

Making simulation part of your content system

Roles and checkpoints (who does what, when)

Successful integration requires clear roles and systematic checkpoints. Here’s a basic framework:

Content creator: Drafts initial copy, runs first synthetic test, makes initial edits

Content lead: Reviews sub-80 scores, decides on iteration vs escalation

Product marketing: Provides persona data, validates synthetic feedback against market knowledge

Performance marketing: Tracks correlation between scores and actual performance

Build checkpoints into your workflow:

Pre-design: Test core messages before visual investment

Pre-production: Validate final copy before video/asset creation

Pre-launch: Final synthetic review before campaign activation

Post-launch: Compare synthetic predictions to actual results

Automation and human checks

The goal isn’t to automate everything but to create intelligent systems where people and AI work together.

So, here’s an orchestration pattern that balances both.

Automated synthetic testing runs on all copy variations, flagging anything below threshold. Human review focuses on edge cases: scores between 75-85, conflicting feedback from different personas, or copy for new market segments where synthetic accuracy hasn’t been validated.

Think of it as building a content assembly line where AI handles the repetitive quality checks while humans make nuanced decisions. The synthetic personas become your first readers, catching obvious problems before human time is invested in refinement.

This approach embodies the principle of designing intelligent systems where clear roles and reliable workflows produce consistent, on-brand content at scale. We’re not replacing human judgment but focusing it where it matters most.

A small counterintuitive rule to keep you honest

Here’s the rule that keeps me honest: reject anything that scores above 95.

When synthetic personas love everything about your copy, something’s wrong. Either your persona isn’t properly calibrated, or your copy is so generic that it offends no one and excites no one. Real customers have preferences, pet peeves, and particular ways they want to be spoken to. If your synthetic persona doesn’t push back on anything, you haven’t built a real simulation, but an echo chamber.

The best copy lives in that 80-90 range: strong enough to engage, specific enough to occasionally bristle against personal preferences, clear enough to communicate value without being so broad it becomes meaningless.

This practice of customer simulation before writing has shifted something fundamental in how I work. The blank page isn’t blank anymore. Before I write, I rehearse. Before I polish, I test. Before I ship, I know where the friction lives.

The question isn’t whether AI can simulate your customers, because the scientific evidence confirms it can. The question is whether you’ll build the discipline to use simulation as a tool for rigour rather than a shortcut to consensus.

Start small. Take your next headline and run it through a synthetic persona built from your worst customer review. Listen to what comes back. That discomfort you feel? That’s the sound of your copy getting better before it meets the world.

If you’d like to explore building a simulation workflow for your next campaign, get in touch to discuss your content or product marketing needs.

Thanks for reading Cowritten.ai! Subscribe for free to receive new posts and support my work.